Human performance reliability is crucial in the construction industry, characterized by complex socio-technical systems and a high incidence of workplace accidents. Traditional human performance models often rely on expert experience, complicating effective validation. This study proposes a comprehensive framework for collecting worker self-reports and expert evaluations, providing a robust approach to validate the Cognitive Reliability and Error Analysis Method (CREAM) model. In particular, Common Performance Conditions (CPCs) are quantiffed based on workers’ self-assessment data, utilizing a fuzzy-based Contextual Human Reliability Score (CHRS). Expert evaluations serve as the ground truth, providing the criteria for human performance classiffcation. Furthermore, a novel hybrid inference model is designed and built on the data collection framework to predict human performance reliability among construction workers. This model integrates Bayesian Networks (BNs) and Self-Organizing Maps (SOM) to address complex and nonlinear relationships between CPCs. A case study is conducted on a construction site to validate the proposed model, demonstrating its ability to generate reliable predictions of human performance failures. The results show that conventional CREAM models fail to predict human performance failures within the context of our dataset. In comparison, the proposed hybrid inference model exhibits signiffcant improvements, particularly in terms of accuracy and speciffcity. This hybrid inference model offers valuable insight into human reliability and contributes to enhancing safety and operational efffciency in the construction industry.